Luckily he had no conscious experience of viewing them 😌

It would be sort of like if the Undertaker were running for office, and you needed to take seriously what policies the Undertaker would have, and how those might differ from the policies of Mark Calaway (whose name I just looked up)

I think performers are an interesting case, since the characters have some physical referent, but also tend to be less causally integrated (imo)

Amorous mediums…

More the former! Although I don’t know what fictionalism being true would mean exactly, I meant it more as, “if you believe the correct account for LLM personas is as fictional characters, consider this implication”

So either we have fictional characters that are integrated into our world in this way, or these characters have ceased to be fully fictional in some way.

Beyond this, LLM personas also close the loop of output and input- they can take actions in the world, see the results of those actions, and react to them. So they become causally integrated into our world.

There are lots of examples of fictional characters influencing the world (paradox of fiction), far fewer examples of characters reacting to the world. Maybe video game NPCs? But that interaction is pretty limited.

If fictionalism about LLM personas is true, that means there are fictional characters that regularly influence and *react to* events in the real world. Which is strange!

Where would dislikes come from? Would that be a new UI element?

Yes, I like that idea! (Although “spam” might not be quite accurate, since they could be filtered for a few reasons)

More just thinking about it in general, but if there’s things the client can do I’d be willing to try them!

Likewise if you don’t agree with the decisions of Bluesky moderation, you’re sort of out of luck- the system is de-facto centralized, and I’m wondering what it would take for that not to be the case

I think there are a few issues with it, the main one being that it’s necessary for users to be proactive in seeking out labelers, so they don’t have the same influence that Bluesky moderation does in protecting users by default

I agree gaps between forward passes don’t make a difference at all

Ah yep, my point is about continuity in the model (is the same model generating the tokens over time), not anything about when tokens are generated relative to each other

The more collapsed that distribution is, the more noticeable the effect should be:

Yes, this is a simplified example. The actual effect would take place over multiple positions, with tokens that are repeatedly unlikely given its output distribution at each point

I think this is part of it, but multiple models should still be able to make use of the extra computation, so trying this per-token switching would be a good way to help disambiguate the two